China proposes pre-approval of AI models before export as US bars DeepSeek from civilian government procurement

Two major AI powers are tightening their grip on Chinese AI, each in its own way. China is considering controlling what leaves the country. The US has already shut the door on DeepSeek in parts of the public sector.

China's proposal has not been adopted

China's Ministry of Commerce (MOFCOM) has proposed that the most advanced AI models, so-called frontier models, must be pre-approved before they are shared internationally. Under the proposal, open-source technologies, meaning models whose code is freely available, would receive automatic approval.

According to Reuters (July 7, 2026), the ministry has held closed-door talks with Alibaba, ByteDance and Z.ai about restrictions on DeepSeek, Qwen and GLM-5.2, as well as on frontier models that have not yet been released. The information comes from three anonymous sources, and the companies declined to comment.

Legal experts affiliated with the Supreme People's Court outlined a three-tier system in May 2026:

- Basic open-source models must be registered. - Advanced technologies must undergo security assessments. - Frontier models must be kept confidential or used only in China.

These are proposals and discussions. The rules have not yet been adopted.

US: The order already applies

In the US, the DHS, the Department of Homeland Security, has issued an order against DeepSeek. According to another source, it covers both DeepSeek as a company and the company's models in connection with civilian government procurement. It is unclear how the order applies to fine-tuned and derivative models, meaning models built on top of DeepSeek's.

The order thus applies to government procurement. The sources do not say that it affects private companies or ordinary users.

New DeepSeek model

At the same time, DeepSeek has released DeepSeek-V4.1-Flash-1017 under the MIT license, a very permissive license. The model is a sparse mixture-of-experts model with 552 billion parameters and can accept images as input. Sparse mixture-of-experts means that only part of the model is in use at a time.

According to the AI Weekly newsletter, the model leads on the DeepSWE benchmark, but this has not been independently verified.